[Paper Review] The Use of Self Organizing Map Method and Feature Selection in Image Database Classification System

Dian Pratiwi|arXiv (Cornell University)|Jun 1, 2012

Image Retrieval and Classification Techniques9 references19 citations

TL;DR

This paper proposes a hybrid image classification system using Self Organizing Maps (SOM) combined with feature selection via PCA and LSA. It processes 250 color images through grayscale conversion, color histogram extraction, and dimensionality reduction before clustering into five classes; PCA-based feature selection with 100 components achieved 88% accuracy, outperforming LSA's 74%.

ABSTRACT

This paper presents a technique in classifying the images into a number of classes or clusters desired by means of Self Organizing Map (SOM) Artificial Neural Network method. A number of 250 color images to be classified as previously done some processing, such as RGB to grayscale color conversion, color histogram, feature vector selection, and then classifying by the SOM Feature vector selection in this paper will use two methods, namely by PCA (Principal Component Analysis) and LSA (Latent Semantic Analysis) in which each of these methods would have taken the characteristic vector of 50, 100, and 150 from 256 initial feature vector into the process of color histogram. Then the selection will be processed into the SOM network to be classified into five classes using a learning rate of 0.5 and calculated accuracy. Classification of some of the test results showed that the highest percentage of accuracy obtained when using PCA and the selection of 100 feature vector that is equal to 88%, compared to when using LSA selection that only 74%. Thus it can be concluded that the method fits the PCA feature selection methods are applied in conjunction with SOM and has an accuracy rate better than the LSA feature selection methods. Keywords: Color Histogram, Feature Selection, LSA, PCA, SOM.

Motivation & Objective

• To improve image database classification accuracy using unsupervised learning techniques.
• To evaluate the impact of feature selection methods—PCA and LSA—on classification performance.
• To determine the optimal number of features (50, 100, 150) for effective clustering using SOM.
• To compare the performance of PCA and LSA in reducing dimensionality while preserving discriminative image features.
• To validate the effectiveness of combining SOM with feature selection for image clustering tasks.

Proposed method

• Convert input color images to grayscale to standardize image representation.
• Extract color histograms as initial feature vectors, resulting in 256-dimensional feature space per image.
• Apply Principal Component Analysis (PCA) and Latent Semantic Analysis (LSA) for feature selection to reduce dimensionality to 50, 100, and 150 components.
• Train a Self Organizing Map (SOM) neural network with a learning rate of 0.5 to classify images into five predefined clusters.
• Use the SOM's competitive learning mechanism to map high-dimensional feature vectors onto a 2D lattice for clustering.
• Evaluate classification accuracy by comparing predicted cluster assignments with ground truth labels.

Experimental results

Research questions

• RQ1 How does PCA-based feature selection compare to LSA in improving image classification accuracy using SOM?
• RQ2 What is the optimal number of features (50, 100, 150) for maximizing classification accuracy in this image clustering system?
• RQ3 Does dimensionality reduction via PCA or LSA preserve sufficient discriminative information for effective image clustering?
• RQ4 Can the combination of SOM and feature selection achieve high accuracy in unsupervised image database classification?
• RQ5 What is the performance difference between PCA and LSA when applied to color histogram features in image clustering?

Key findings

• The highest classification accuracy of 88% was achieved when using PCA with 100 selected features.
• LSA-based feature selection yielded a lower accuracy of 74% under the same clustering conditions.
• PCA consistently outperformed LSA in preserving relevant image features for clustering, especially at 100-component selection.
• The use of 150 features did not significantly improve accuracy over 100 features, suggesting diminishing returns beyond a certain threshold.
• The SOM network successfully grouped images into five distinct clusters when combined with PCA-based feature selection.
• The results indicate that PCA is a more effective feature selection method than LSA for this image classification task using SOM.